Signal-based classification of cementitious materials using machine learning techniques

The material properties of cementitious mixtures are critically influenced by their mixing ratios and curing periods. However, existing nondestructive testing methods based on elastic waves primarily provide information on stiffness characteristics, without offering insights into mix design or curing duration. This study proposes a machine learning (ML)-based approach for estimating the mixing ratios and curing periods of cementitious materials using time-series shear wave signals obtained from bender elements. Three ML models—one-dimensional convolutional neural network, InceptionTime, and Random Convolutional Kernel Transform—are applied to classify 16 cementitious mixtures. The signal patterns from the bender elements are recorded for four cementitious mixtures with different mixing ratios over four curing periods ranging from 1 to 28 days. Using data from a shear wave measurement system, the models estimate 16 unique classes of mixtures. Model performance is evaluated across training, validation, and test sets. All models generalize well to unseen data, achieving an accuracy greater than 0.98. The InceptionTime model notably achieves the highest accuracy at 0.992. In addition, the confusion matrix analysis confirms that the models produce very few misclassifications, indicating strong predictive performance. Overall, the one-dimensional convolutional neural network offers the best balance between accuracy and computational efficiency. These findings suggest that machine learning techniques for time series analysis can estimate the mixing ratios and curing periods of cementitious materials from signal patterns effectively.